lumin.nn.callbacks package¶
Submodules¶
lumin.nn.callbacks.callback module¶
-
class
lumin.nn.callbacks.callback.
Callback
[source]¶ Bases:
lumin.nn.callbacks.abs_callback.AbsCallback
Base callback class from which other callbacks should inherit.
-
set_model
(model)[source]¶ Sets the callback’s model in order to allow the callback to access and adjust model parameters
- Parameters
model (
AbsModel
) – model to refer to during training- Return type
None
-
set_plot_settings
(plot_settings)[source]¶ Sets the plot settings for any plots produced by the callback
- Parameters
plot_settings (
PlotSettings
) – PlotSettings class- Return type
None
-
lumin.nn.callbacks.cyclic_callbacks module¶
-
class
lumin.nn.callbacks.cyclic_callbacks.
AbsCyclicCallback
(interp, param_range, cycle_mult=1, decrease_param=False, scale=1, cycle_save=False)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Abstract class for callbacks affecting lr or mom
- Parameters
interp (
str
) – string representation of interpolation function. Either ‘linear’ or ‘cosine’.param_range (
Tuple
[float
,float
]) – minimum and maximum values for parametercycle_mult (
int
) – multiplicative factor for adjusting the cycle length after each cycle. E.g cycle_mult=1 keeps the same cycle length, cycle_mult=2 doubles the cycle length after each cycle.decrease_param (
bool
) – whether to begin by decreasing the parameter, otherwise begin by increasing itscale (
int
) – multiplicative factor for setting the initial number of epochs per cycle. E.g scale=1 means 1 epoch per cycle, scale=5 means 5 epochs per cycle.cycle_save (
bool
) – if true will save a copy of the model at the end of each cycle. Used for building ensembles from single trainings (e.g. snapshot ensembles)nb – number of minibatches (iterations) to expect per epoch
-
on_batch_begin
()[source]¶ Computes the new value for the optimiser parameter and passes it to _set_param method
- Return type
None
-
class
lumin.nn.callbacks.cyclic_callbacks.
CycleLR
(lr_range, interp='cosine', cycle_mult=1, decrease_param='auto', scale=1, cycle_save=False)[source]¶ Bases:
lumin.nn.callbacks.cyclic_callbacks.AbsCyclicCallback
Callback to cycle learning rate during training according to either: cosine interpolation for SGDR https://arxiv.org/abs/1608.03983 or linear interpolation for Smith cycling https://arxiv.org/abs/1506.01186
- Parameters
lr_range (
Tuple
[float
,float
]) – tuple of initial and final LRsinterp (
str
) – ‘cosine’ or ‘linear’ interpolationcycle_mult (
int
) – Multiplicative constant for altering the cycle length after each complete cycledecrease_param (
Union
[str
,bool
]) – whether to increase or decrease the LR (effectively reverses lr_range order), ‘auto’ selects according to interpscale (
int
) – Multiplicative constant for altering the length of a cycle. 1 corresponds to one cycle = one epochcycle_save (
bool
) – if true will save a copy of the model at the end of each cycle. Used for building ensembles from single trainings (e.g. snapshot ensembles)nb – Number of batches in a epoch
- Examples::
>>> cosine_lr = CycleLR(lr_range=(0, 2e-3), cycle_mult=2, scale=1, ... interp='cosine', nb=100) >>> >>> cyclical_lr = CycleLR(lr_range=(2e-4, 2e-3), cycle_mult=1, scale=5, interp='linear', nb=100)
-
class
lumin.nn.callbacks.cyclic_callbacks.
CycleMom
(mom_range, interp='cosine', cycle_mult=1, decrease_param='auto', scale=1, cycle_save=False)[source]¶ Bases:
lumin.nn.callbacks.cyclic_callbacks.AbsCyclicCallback
Callback to cycle momentum (beta 1) during training according to either: cosine interpolation for SGDR https://arxiv.org/abs/1608.03983 or linear interpolation for Smith cycling https://arxiv.org/abs/1506.01186 By default is set to evolve in opposite direction to learning rate, a la https://arxiv.org/abs/1803.09820
- Parameters
mom_range (
Tuple
[float
,float
]) – tuple of initial and final momentainterp (
str
) – ‘cosine’ or ‘linear’ interpolationcycle_mult (
int
) – Multiplicative constant for altering the cycle length after each complete cycledecrease_param (
Union
[str
,bool
]) – whether to increase or decrease the momentum (effectively reverses mom_range order), ‘auto’ selects according to interpscale (
int
) – Multiplicative constant for altering the length of a cycle. 1 corresponds to one cycle = one epochcycle_save (
bool
) – if true will save a copy of the model at the end of each cycle. Used for building ensembles from single trainings (e.g. snapshot ensembles)nb – Number of batches in a epoch
- Examples::
>>> cyclical_mom = CycleMom(mom_range=(0.85 0.95), cycle_mult=1, ... scale=5, interp='linear', nb=100)
-
class
lumin.nn.callbacks.cyclic_callbacks.
OneCycle
(lengths, lr_range, mom_range=(0.85, 0.95), interp='cosine')[source]¶ Bases:
lumin.nn.callbacks.cyclic_callbacks.AbsCyclicCallback
Callback implementing Smith 1-cycle evolution for lr and momentum (beta_1) https://arxiv.org/abs/1803.09820 Default interpolation uses fastai-style cosine function. Automatically triggers early stopping on cycle completion.
- Parameters
lengths (
Tuple
[int
,int
]) – tuple of number of epochs in first and second stages of cyclelr_range (
List
[float
]) – list of initial and max LRs and optionally a final LR. If only two LRs supplied, then final LR will be zero.mom_range (
Tuple
[float
,float
]) – tuple of initial and final momentainterp (
str
) – ‘cosine’ or ‘linear’ interpolationnb – Number of batches in a epoch
- Examples::
>>> onecycle = OneCycle(lengths=(15, 30), lr_range=[1e-4, 1e-2], ... mom_range=(0.85, 0.95), interp='cosine', nb=100)
lumin.nn.callbacks.data_callbacks module¶
-
class
lumin.nn.callbacks.data_callbacks.
BinaryLabelSmooth
(coefs=0)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Callback for applying label smoothing to binary classes, based on https://arxiv.org/abs/1512.00567 Applies smoothing during both training.
- Parameters
coefs (
Union
[float
,Tuple
[float
,float
]]) – Smoothing coefficients: 0->coef[0] 1->1-coef[1]. if passed float, coef[0]=coef[1]
- Examples::
>>> lbl_smooth = BinaryLabelSmooth(0.1) >>> >>> lbl_smooth = BinaryLabelSmooth((0.1, 0.02))
-
class
lumin.nn.callbacks.data_callbacks.
BootstrapResample
(n_folds, bag_each_time=False, reweight=True)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Callback for bootstrap sampling new training datasets from original training data during (ensemble) training.
- Parameters
n_folds (
int
) – the number of folds present in trainingFoldYielder
bag_each_time (
bool
) – whether to sample a new set for each sub-epoch or to use the same sample each timereweight (
bool
) – whether to reweight the sampleed data to mathch the weight sum (per class) of the original data
- Examples::
>>> bs_resample BootstrapResample(n_folds=len(train_fy))
-
class
lumin.nn.callbacks.data_callbacks.
ParametrisedPrediction
(feats, param_feat, param_val)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Callback for running predictions for a parametersied network (https://arxiv.org/abs/1601.07913); one which has been trained using one of more inputs which represent e.g. different hypotheses for the classes such as an unknown mass of some new particle. In such a scenario, multiple signal datasets could be used for training, with background receiving a random mass. During prediction one then needs to set these parametrisation features all to the same values to evaluat the model’s response for that hypothesis. This callback can be passed to the predict method of the model/ensemble to adjust the parametrisation features to the desired values.
- Parameters
feats (
List
[str
]) – list of feature names used during training (in the same order)param_feat (
Union
[List
[str
],str
]) – the feature name which is to be adjusted, or a list of features to adjustparam_val (
Union
[List
[float
],float
]) – the value to which to set the paramertisation feature, of the list of values to set the parameterisation features to
- Examples::
>>> mass_param = ParametrisedPrediction(train_feats, 'res_mass', 300) >>> model.predict(fold_yeilder, pred_name=f'pred_mass_300', callbacks=[mass_param]) >>> >>> mass_param = ParametrisedPrediction(train_feats, 'res_mass', 300) >>> spin_param = ParametrisedPrediction(train_feats, 'spin', 1) >>> model.predict(fold_yeilder, pred_name=f'pred_mass_300', callbacks=[mass_param, spin_param])
lumin.nn.callbacks.loss_callbacks module¶
-
class
lumin.nn.callbacks.loss_callbacks.
GradClip
(clip, clip_norm=True)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Callback for clipping gradients by norm or value.
- Parameters
clip (
float
) – value to clip atclip_norm (
bool
) – whether to clip according to norm (torch.nn.utils.clip_grad_norm_) or value (torch.nn.utils.clip_grad_value_)
- Examples::
>>> grad_clip = GradClip(1e-5)
lumin.nn.callbacks.lsuv_init module¶
This file contains code modfied from https://github.com/ducha-aiki/LSUV-pytorch which is made available under the following BSD 2-Clause “Simplified” Licence: Copyright (C) 2017, Dmytro Mishkin All rights reserved.
Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: 1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
The Apache Licence 2.0 underwhich the majority of the rest of LUMIN is distributed does not apply to the code within this file.
-
class
lumin.nn.callbacks.lsuv_init.
LsuvInit
(needed_std=1.0, std_tol=0.1, max_attempts=10, do_orthonorm=True, verbose=False)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Applies Layer-Sequential Unit-Variance (LSUV) initialisation to model, as per Mishkin & Matas 2016 https://arxiv.org/abs/1511.06422. When training begins for the first time, Conv1D, Conv2D, Conv3D, and Linear modules in the model will be LSUV initialised using the BatchYielder inputs. This involves initialising the weights with orthonormal matirces and then iteratively scaling them such that the stadndar deviation of the layer outputs is equal to a desired value, within some tolerance.
- Parameters
needed_std (
float
) – desired standard deviation of layer outputsstd_tol (
float
) – tolerance for matching standard deviation with targetmax_attempts (
int
) – number of times to attempt weight scaling per layerdo_orthonorm (
bool
) – whether to apply orthonormal initialisation first, or rescale the exisiting valuesverbose (
bool
) – whether to print out details of the rescaling
- Example::
>>> lsuv = LsuvInit() >>> >>> lsuv = LsuvInit(verbose=True) >>> >>> lsuv = LsuvInit(needed_std=0.5, std_tol=0.01, max_attempts=100, do_orthonorm=True)
lumin.nn.callbacks.model_callbacks module¶
-
class
lumin.nn.callbacks.model_callbacks.
SWA
(start_epoch, renewal_period=None, update_on_cycle_end=None, verbose=False)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Callback providing Stochastic Weight Averaging based on (https://arxiv.org/abs/1803.05407) This adapted version allows the tracking of a pair of average models in order to avoid having to hardcode a specific start point for averaging:
Model average #0 will begin to be tracked start_epoch epochs/cycles after training begins.
cycle_since_replacement is set to 1
Renewal_period epochs/cycles later, a second average #1 will be tracked.
At the next renewal period, the performance of #0 and #1 will be compared on data contained in val_fold.
- If #0 is better than #1:
#1 is replaced by a copy of the current model
cycle_since_replacement is increased by 1
renewal_period is multiplied by cycle_since_replacement
- Else:
#0 is replaced by #1
#1 is replaced by a copy of the current model
cycle_since_replacement is set to 1
renewal_period is set back to its original value
Additonally, will optionally (default True) lock-in to any cyclical callbacks to only update at the end of a cycle.
- Parameters
start_epoch (
int
) – epoch/cycle to begin averagingrenewal_period (
Optional
[int
]) – How often to check performance of averages, and renew tracking of least performant. If None, will not track a second average.update_on_cycle_end (
Optional
[bool
]) – Whether to lock in to the cyclic callback and only update at the end of a cycle. Default yes, if cyclic callback present.verbose (
bool
) – Whether to print out update information for testing and operation confirmation
- Examples::
>>> swa = SWA(start_epoch=5, renewal_period=5)
lumin.nn.callbacks.monitors module¶
-
class
lumin.nn.callbacks.monitors.
EarlyStopping
(patience, loss_is_meaned=True)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Tracks validation loss during training and terminates training if loss doesn’t decrease after patience number of epochs. Losses are assumed to be averaged and will be re-averaged over the epoch unless loss_is_meaned is false.
- Parameters
patience (
int
) – number of epochs to wait without improvement before stopping trainingloss_is_meaned (
bool
) – if the batch loss value has been averaged over the number of elements in the batch, this should be true; average loss will be computed over all elements in batch. If the batch loss is not an average value, then the average will be computed over the number of batches.
-
class
lumin.nn.callbacks.monitors.
SaveBest
(auto_reload=True, loss_is_meaned=True)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Tracks validation loss during training and automatically saves a copy of the weights to indicated file whenever validation loss decreases. Losses are assumed to be averaged and will be re-averaged over the epoch unless loss_is_meaned is false.
- Parameters
auto_reload (
bool
) – if true, will automatically reload the best model at the end of trainingloss_is_meaned (
bool
) – if the batch loss value has been averaged over the number of elements in the batch, this should be true; average loss will be computed over all elements in batch. If the batch loss is not an average value, then the average will be computed over the number of batches.
-
class
lumin.nn.callbacks.monitors.
MetricLogger
(show_plots=False, extra_detail=True, loss_is_meaned=True)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Provides live feedback during training showing a variety of metrics to help highlight problems or test hyper-parameters without completing a full training. If show_plots is false, will instead print training and validation losses at the end of each epoch. The full history is available as a dictionary by calling
get_loss_history()
.- Parameters
loss_names – List of names of losses which will be passed to the logger in the order in which they will be passed. By convention the first name will be used as the training loss when computing the ratio of training to validation losses
n_folds – Number of folds present in the training data. The logger assumes that one of these folds is for validation, and so 1 training epoch = (n_fold-1) folds.
autolog_scale – Whether to automatically change the scale of the y-axis for loss to logarithmic when the current loss drops below one 50th of its starting value
extra_detail (
bool
) – Whether to include extra detail plots (loss velocity and training validation ratio), slight slower but potentially useful.
-
get_loss_history
()[source]¶ Get the current history of losses
- Returns
ordered dictionary (training first, validations subsequent) mapping loss names to lists of loss values
- Return type
history
lumin.nn.callbacks.opt_callbacks module¶
-
class
lumin.nn.callbacks.opt_callbacks.
LRFinder
(lr_bounds=[1e-07, 10], nb=None)[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Callback class for Smith learning-rate range test (https://arxiv.org/abs/1803.09820)
- Parameters
nb (
Optional
[int
]) – number of batches in a epochlr_bounds (
Tuple
[float
,float
]) – tuple of initial and final LR
lumin.nn.callbacks.pred_handlers module¶
-
class
lumin.nn.callbacks.pred_handlers.
PredHandler
[source]¶ Bases:
lumin.nn.callbacks.callback.Callback
Default callback for predictions. Collects predictions over batches and returns them as stacked array